Related Experiment Video
Updated: Jul 9, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
GCLR: A self-supervised representation learning pretext task for glomerular filtration barrier segmentation in TEM
Guoyu Lin1, Zhentai Zhang1, Kaixing Long1
1School of Biomedical Engineering, Southern Medical University, Guangzhou, 510515, China; Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou, 510515, China; Guangdong Province Engineering Laboratory for Medical Imaging and Diagnostic Technology, Southern Medical University, Guangzhou, 510515, China.
This study introduces GCLR, a novel self-supervised learning method for segmenting the glomerular filtration barrier (GFB) in TEM images. GCLR improves renal disease diagnosis by effectively utilizing unlabeled data, achieving state-of-the-art segmentation results.
Area of Science:
- Medical Imaging
- Computational Pathology
- Biomedical Engineering
Background:
- Accurate segmentation of the glomerular filtration barrier (GFB) in transmission electron microscopy (TEM) images is crucial for diagnosing renal diseases.
- Manual annotation of TEM images is time-consuming, limiting the availability of training data for deep learning models.
- Self-supervised representation learning (SSRL) offers a promising approach to leverage large unlabeled datasets.
Purpose of the Study:
- To develop an automated method for segmenting the three substructures of the GFB in TEM images.
- To address the challenge of limited annotated data by employing SSRL.
- To introduce GCLR, a novel hybrid pixel-level pretext task for GFB segmentation.
Main Methods:
- Developed GCLR, a hybrid self-supervised learning framework integrating global clustering (GC) and local restoration (LR) pretext tasks.
- Utilized 18,928 unlabeled glomerular TEM images for self-supervised pre-training.
- Fine-tuned the model on 311 labeled images for GFB substructure segmentation.
Main Results:
- GCLR achieved state-of-the-art segmentation performance for all three GFB substructures, with Dice similarity coefficients of 86.56 ± 0.16%, 75.56 ± 0.36%, and 79.41 ± 0.16%.
- Outperformed other representative self-supervised pretext tasks in GFB segmentation.
- Demonstrated superior performance compared to fully-supervised pre-training methods on public datasets (MitoEM, COCO, ImageNet) with reduced data and time.
Conclusions:
- GCLR effectively utilizes unlabeled TEM data for accurate GFB segmentation, mitigating annotation scarcity.
- The proposed method significantly advances automated analysis of renal pathology through improved TEM image segmentation.
- GCLR shows potential for enhancing diagnostic accuracy and efficiency in renal disease assessment.
